An explainable framework for partial discharge detection in power cables through the integration of rough set theory and deep learning

• Proposes a hybrid framework combining machine-driven feature extraction and a knowledge-based classifier for PD identification in power cables. • • Deep features are extracted using a fully-connected autoencoder enhanced by morphological denoising. • Integrates rough set theory with a novel feature analysis algorithm for interpretable, rule based classification. • Enables transparent decision-making by tracking the reasoning behind each classification output Recognition of partial discharge (PD) in power cables' insulation during the early stage of PD progression can trigger a set of preventive actions which significantly reduce power systems breakdowns and service interruptions. Thus, developing effective identification tools is of great importance to prevent such unwanted breakdowns. This paper proposes a hybrid framework based on a knowledge-based classifier and machine-driven feature extraction for PD identification in power cables. The classifier in this framework relies on rough set theory (RST) and a novel feature analysis algorithm (FAA) for PD and noise patterns discrimination. Unlike the ANN-based classifiers (black-box classifiers) that might generate untruthful outputs in case of the input sample and estimated function inconsistency, the RST as the prime classifier, will prevent any classification in the case of non-compliance with the classification rules and activate the FAA for classification using shallow and deep analyses. As a result, by overcoming black-box classifiers' limitations, a meager false detection rate and highly reliable outputs guided by the proposed white-box classifier can be attributed to this framework's main advantages. Also, to enhance classification performance, an improved denoising technique based on morphological filters incorporates a fully-connected autoencoder (FCA) to provide an abstract version of signals (synthetic features) for the classifier. Finally, the effectiveness of the proposed framework is demonstrated through extensive simulations and experimental validation, achieving higher accuracy than existing state-of-the-art approaches (99.42 %).